Face super resolution based on ESRGAN (https://github.com/xinntao/BasicSR)
INPUT & AFTER-SR & GROUND TRUTH
-
Run python gen_lr_imgs.py to get the face imgs with low resolution and pool qualities
-
Set the dir in train.py
hr_path: The path list of imgs with high resolution.
lr_path: The path of imgs with low resolution.
- Run python train.py
- Dlib alignment shape_predictor_68_face_landmarks.dat
(https://pan.baidu.com/s/19Y-AYnXs6ubIh4vlkyvqbQ)
(https://drive.google.com/open?id=1u3h3nX5f_w-HJV8Nd1zwqc3uTnVja5Ol)
- Generator weights
90000_G.pth
(https://pan.baidu.com/s/14ITkNz_t0E7hRv0-tTAjhA)
(https://drive.google.com/open?id=1CZkLZPtbJepgksCM93MvsY7NgqnEZSvk)
90000_G.pth (The last activation in G is linear, clearer)
200000_G.pth
(https://pan.baidu.com/s/1Osge_4JjPyvG5Xfnbe9KVA)
(https://drive.google.com/open?id=1B6BQu5Qk8eIu8MGTWJHnJxaxY1zCqQEt)
200000_G.pth (The last activation in G is tanh)
-
Download 'shape_predictor_68_face_landmarks.dat' and '90000_G.pth'
-
Set 'pretrain_model_G' in test.py
-
RUN python test.py